Our AI couldn't answer a question that sounds obvious: which channel actually makes us money? Not because the model was slow on the uptake, but because nobody had ever put the data somewhere it could actually be joined together. It took us a while to realize the problem wasn't the question, or the AI, it was that the answer was scattered across four places that never talked to each other.
We're telling it exactly as it happened because it's a problem every small business that posts on several channels and bills through a separate system runs into: "how many likes do I have" is easy to check, and "which channel actually brings me customers" almost never is.
The problem: four dashboards, no answer
At AutoBoost we publish content daily on LinkedIn, Threads, Facebook and Instagram, on top of running the blog and billing subscriptions through Stripe. Each piece lives natively in its own dashboard: LinkedIn's panel counts LinkedIn impressions, Meta's panel counts Instagram and Facebook, and Stripe counts revenue without knowing which channel the paying customer originally came from.
If you ask an AI "which channel makes me the most money" with those four dashboards open, the only thing it can do is guess from disconnected fragments, at best comparing how many likes each channel got. It can't connect "this visit came from Threads" with "this customer paid 400 euros," because that relationship doesn't exist anywhere. The AI doesn't fail because of the model. It fails because the data it needs to answer doesn't exist in one unified place.
Why it happened: the data lived in four places that didn't talk to each other
Before touching any AI, we did an inventory of where every piece of information actually lived:
- Social channels: impressions, clicks and saves, each network with its own API and its own field names (which also change without warning, as happened to us with Meta's metrics).
- Website: visits, reading time and which channel each session came from, tracked with our own tracker.
- Leads: who filled in the contact form or wrote in on WhatsApp, and which article they came from.
- Revenue: who paid, how much and when, inside Stripe's ledger.
Four systems, four different identities for the same person. None of them spoke to the other three. And without that conversation, neither a marketing director nor an AI can answer "which channel makes me money": they can only answer "which channel has the most impressions," which is a very different question, and a much less useful one for the business.
What we built: a database, not a new tool
The solution wasn't buying marketing attribution software (it exists, and it's expensive). It was building an internal app that centralizes 3 social channels into a single database, keeping 30 days of daily metrics so we can see a trend, not just today's snapshot. It's 4 apps (blog, social, leads, dashboard) sharing the same middleware and the same database, instead of four separate systems with manual exports in between.
The detail we liked most about the whole project: revenue didn't need any new tooling at all. It was already there, in Stripe's ledger, waiting for someone to connect it to the rest. The real work wasn't "build something to measure revenue," it was linking what already existed: the slug of the article that brought the visit, the channel that visit came from, and the ID of the customer who ended up paying, all under the same reference.
That's, quite literally, "sort the data first, then apply the AI": the hard part, and the part that's actually worth money, isn't the AI, it's getting the data into a state where a real business question has an answer.
What the AI can ask that data now
With all four sources under the same database, the question that used to be impossible now has a direct answer: which channel brought the visit, which visit turned into a lead, and which lead ended up paying. The AI (or any dashboard, with or without AI) no longer opines about likes: it joins real identities and answers with a full funnel, not four disconnected numbers you can't add together.
It's the exact same pattern we used in a case where we built an AI data platform for a pharmacy group: the AI didn't start reasoning properly about sales until the data from several branches was unified in one place. Here the business is different (marketing instead of pharmacy sales), but the lesson is identical.
The checklist before you put an AI on top of your dashboards
If you're thinking about asking an AI which channel, product or campaign actually makes you money, check this before you write the first prompt:
| Question | If the answer is "no" |
|---|---|
| Do your channels and your revenue share a common identifier (a customer, a lead, an order)? | The AI can only compare disconnected numbers, never cause and effect |
| Does your revenue already live in a reliable system (billing, bank, Stripe)? | Don't build a new dashboard for something that already exists, connect it instead |
| Can you answer "which channel makes money" without opening more than one tab? | That's the real pending work, and it comes before any AI |
| Do you keep history, not just today's snapshot? | Without a trend, every decision is based on a single day, good or bad |
If any answer is "no," that's the project to do first. It's not a boring prerequisite before "the AI part": it's the part that makes the AI, once it arrives, say something true instead of a nice-sounding opinion about likes.
How we do this at AutoBoost
This is the same approach we apply on client projects: before promising an assistant or an AI agent, we check whether the data that agent needs actually lives somewhere it can be fed properly. Often the most profitable project isn't "add AI," it's getting the data in shape, something we also cover in detail in our services.
If your business has the same problem (scattered dashboards, none of them tied to revenue), tell us what data you handle today and how many different places it lives in. We can help you see it clearly. Talk to us.
